[CE] Add CE for Hybrid Parallism (#7817)

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N1C8/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-SD2-Stage1-mbs8-acc2.sh → tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N1C4/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-SD2-Stage1-mbs8-acc2.sh

@@ -19,7 +19,7 @@ pp_degree=1
 bs_item=32
 fp_item=bf16
 run_mode=MP2-SP2-SD2-Stage1-mbs8-acc2
-device_num=N1C8
+device_num=N1C4
 max_iter=50000
 sharding=stage1
 sharding_degree=2

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N1C8/gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-SD2-Stage1-mbs8-acc2.sh → tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N1C4/gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-SD2-Stage1-mbs8-acc2.sh

@@ -19,7 +19,7 @@ pp_degree=1
 bs_item=32
 fp_item=bf16
 run_mode=MP2-SP2-SD2-Stage1-mbs8-acc2
-device_num=N1C8
+device_num=N1C4
 max_iter=100
 sharding=stage1
 sharding_degree=2

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N1C8/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-PP2-SD2-Stage1-mbs8-acc2.sh

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-model_item=gpt-345m_seqlen1024_pretrain
+model_item=CE_gpt-345m_seqlen1024_pretrain
 dp_degree=1
 mp_degree=2
 pp_degree=2

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-PP4-VPP2-DP2-mbs2-acc8.sh

@@ -0,0 +1,41 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+model_item=CE_gpt-345m_seqlen1024_pretrain
+dp_degree=2
+mp_degree=2
+pp_degree=4
+bs_item=32
+fp_item=bf16
+run_mode=MP2-PP4-VPP2-DP2-mbs2-acc8
+device_num=N2C16
+max_iter=50000
+sharding=False
+sharding_degree=1
+
+virtual_pp_degree=2
+use_recompute=True
+eval_freq=25
+use_pipeline_parallel=True
+sequence_parallel=False
+
+model=gpt
+micro_bs=2
+acc=8
+seed=1234
+
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
+# run
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh ${model_item} ${fp_item} ${dp_degree} ${mp_degree} ${pp_degree} ${micro_bs} ${bs_item} ${run_mode} ${device_num} \
+${max_iter} ${sharding} ${sharding_degree} ${virtual_pp_degree} ${use_recompute} ${eval_freq} ${use_pipeline_parallel} ${sequence_parallel} ${acc} ${seed} 2>&1;

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-PP4-VPP2-SD2-stage1-mbs2-acc8.sh

@@ -0,0 +1,41 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+model_item=CE_gpt-345m_seqlen1024_pretrain
+dp_degree=1
+mp_degree=2
+pp_degree=4
+bs_item=32
+fp_item=bf16
+run_mode=MP2-PP4-VPP2-SD2-stage1-mbs2-acc8
+device_num=N2C16
+max_iter=50000
+max_iter=50000
+sharding=stage1
+sharding_degree=2
+virtual_pp_degree=2
+use_recompute=True
+eval_freq=25
+use_pipeline_parallel=True
+sequence_parallel=False
+
+model=gpt
+micro_bs=2
+acc=8
+seed=1234
+
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
+# run
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh ${model_item} ${fp_item} ${dp_degree} ${mp_degree} ${pp_degree} ${micro_bs} ${bs_item} ${run_mode} ${device_num} \
+${max_iter} ${sharding} ${sharding_degree} ${virtual_pp_degree} ${use_recompute} ${eval_freq} ${use_pipeline_parallel} ${sequence_parallel} ${acc} ${seed} 2>&1;

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-PP4-VPP2-DP2-mbs2-acc8.sh

@@ -0,0 +1,41 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+model_item=CE_gpt-345m_seqlen1024_pretrain
+dp_degree=2
+mp_degree=2
+pp_degree=4
+bs_item=32
+fp_item=bf16
+run_mode=MP2-SP2-PP4-VPP2-DP2-mbs2-acc8
+device_num=N2C16
+max_iter=50000
+sharding=False
+sharding_degree=1
+
+virtual_pp_degree=2
+use_recompute=True
+eval_freq=25
+use_pipeline_parallel=True
+sequence_parallel=True
+
+model=gpt
+micro_bs=2
+acc=8
+seed=1234
+
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
+# run
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh ${model_item} ${fp_item} ${dp_degree} ${mp_degree} ${pp_degree} ${micro_bs} ${bs_item} ${run_mode} ${device_num} \
+${max_iter} ${sharding} ${sharding_degree} ${virtual_pp_degree} ${use_recompute} ${eval_freq} ${use_pipeline_parallel} ${sequence_parallel} ${acc} ${seed} 2>&1;

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/CE_gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-PP4-VPP2-SD2-stage1-mbs2-acc8.sh

@@ -32,7 +32,6 @@ sequence_parallel=True
 model=gpt
 micro_bs=2
 acc=8
-seed=3589
 
 bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
 # run

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-PP4-VPP2-DP2-mbs2-acc8.sh

@@ -0,0 +1,41 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+model_item=gpt-345m_seqlen1024_pretrain
+dp_degree=2
+mp_degree=2
+pp_degree=4
+bs_item=32
+fp_item=bf16
+run_mode=MP2-PP4-VPP2-DP2-mbs2-acc8
+device_num=N2C16
+max_iter=100
+sharding=False
+sharding_degree=1
+
+virtual_pp_degree=2
+use_recompute=True
+eval_freq=25
+use_pipeline_parallel=True
+sequence_parallel=False
+
+model=gpt
+micro_bs=2
+acc=8
+seed=1234
+
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
+# run
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh ${model_item} ${fp_item} ${dp_degree} ${mp_degree} ${pp_degree} ${micro_bs} ${bs_item} ${run_mode} ${device_num} \
+${max_iter} ${sharding} ${sharding_degree} ${virtual_pp_degree} ${use_recompute} ${eval_freq} ${use_pipeline_parallel} ${sequence_parallel} ${acc} ${seed} 2>&1;

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-PP4-VPP2-SD2-stage1-mbs2-acc8.sh

@@ -0,0 +1,40 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+model_item=gpt-345m_seqlen1024_pretrain
+dp_degree=1
+mp_degree=2
+pp_degree=4
+bs_item=32
+fp_item=bf16
+run_mode=MP2-PP4-VPP2-SD2-stage1-mbs2-acc8
+device_num=N2C16
+max_iter=100
+sharding=stage1
+sharding_degree=2
+virtual_pp_degree=2
+use_recompute=True
+eval_freq=25
+use_pipeline_parallel=True
+sequence_parallel=False
+
+model=gpt
+micro_bs=2
+acc=8
+seed=1234
+
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
+# run
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh ${model_item} ${fp_item} ${dp_degree} ${mp_degree} ${pp_degree} ${micro_bs} ${bs_item} ${run_mode} ${device_num} \
+${max_iter} ${sharding} ${sharding_degree} ${virtual_pp_degree} ${use_recompute} ${eval_freq} ${use_pipeline_parallel} ${sequence_parallel} ${acc} ${seed} 2>&1;

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/N2C16/gpt-345m_seqlen1024_pretrain_bs32_bf16_MP2-SP2-PP4-VPP2-DP2-mbs2-acc8.sh

@@ -0,0 +1,41 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+model_item=gpt-345m_seqlen1024_pretrain
+dp_degree=2
+mp_degree=2
+pp_degree=4
+bs_item=32
+fp_item=bf16
+run_mode=MP2-SP2-PP4-VPP2-DP2-mbs2-acc8
+device_num=N2C16
+max_iter=100
+sharding=False
+sharding_degree=1
+
+virtual_pp_degree=2
+use_recompute=True
+eval_freq=25
+use_pipeline_parallel=True
+sequence_parallel=True
+
+model=gpt
+micro_bs=2
+acc=8
+seed=1234
+
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/prepare.sh
+# run
+bash ./test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh ${model_item} ${fp_item} ${dp_degree} ${mp_degree} ${pp_degree} ${micro_bs} ${bs_item} ${run_mode} ${device_num} \
+${max_iter} ${sharding} ${sharding_degree} ${virtual_pp_degree} ${use_recompute} ${eval_freq} ${use_pipeline_parallel} ${sequence_parallel} ${acc} ${seed} 2>&1;

tests/test_tipc/dygraph/hybrid_parallelism/ce_gpt/benchmark_common/run_benchmark.sh

@@ -25,7 +25,7 @@ function _set_params(){
     micro_batch_size=${6:-"2"}      # (必选) micro_batch_size
     global_batch_size=${7:-"16"}    # （必选）global_batch_size
     run_mode=${8:-"DP"}             # (必选) MP模型并行|DP数据并行|PP流水线并行|混合并行DP1-MP1-PP1|DP2-MP8-PP2|DP1-MP8-PP4|DP4-MP8-PP1
-    device_num=${9:-"N1C1"}         # (必选) 使用的卡数量，N1C1|N1C8|N4C32 （4机32卡）
+    device_num=${9:-"N1C1"}         # (必选) 使用的卡数量，N1C1|N1C8|N1C4|N4C32 （4机32卡）
     profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
     model_repo="PaddleNLP"          # (必选) 模型套件的名字
     speed_unit="tokens/s"         # (必选)速度指标单位
@@ -153,7 +153,8 @@ function _train(){
             run_pretrain.py ${train_cmd}"
         workerlog_id=0
         ;;
-    DP8-mbs2-acc2|SD8-stage1-mbs2-acc2|SD8-stage2-mbs2-acc2|SD8-stage3-mbs2-acc2|PP8-mbs16-acc2|MP2-SD4-stage1-mbs4-acc2|MP2-SP2-PP2-DP2-mbs8-acc2|MP8-mbs16-acc2|MP2-PP2-DP2-mbs8-acc2|MP2-PP2-SD2-Stage1-mbs8-acc2|MP2-SP2-PP2-SD2-Stage1-mbs8-acc2|MP2-SP2-PP4-VPP2-SD2-stage1-mbs2-acc8) echo "run run_mode: ${run_mode}"
+    DP8-mbs2-acc2|SD8-stage1-mbs2-acc2|SD8-stage2-mbs2-acc2|SD8-stage3-mbs2-acc2|PP8-mbs16-acc2|MP8-mbs16-acc2|MP2-PP2-DP2-mbs8-acc2|MP2-PP2-SD2-Stage1-mbs8-acc2|MP2-SD4-stage1-mbs4-acc2|MP2-SP2-PP2-DP2-mbs8-acc2|MP2-SP2-PP2-SD2-Stage1-mbs8-acc2|MP2-PP4-VPP2-DP2-mbs2-acc8|MP2-PP4-VPP2-SD2-stage1-mbs2-acc8|MP2-SP2-PP4-VPP2-DP2-mbs2-acc8|MP2-SP2-PP4-VPP2-SD2-stage1-mbs2-acc8) 
+    echo "run run_mode: ${run_mode}"
         train_cmd="python -m paddle.distributed.launch --log_dir=./mylog --devices=0,1,2,3,4,5,6,7 ${PADDLE_RANK_OPTION}\
             run_pretrain.py ${train_cmd}"
         workerlog_id=0